Its funny how different in size GPUs can be even though they all accomplish the same task.
keenan
@mdesoto Same task, different speed! Very, very roughly speaking, A13 does about 1 trillion operations per second (for the whole chip, including CPU+GPU), whereas a GeForce 3090 RTX does about 36 trillion operations per second. The former has about 8.5 billion transistors; the latter about 28.3 billion. Also, totally different power budget. Not an apples-to-apples comparison at all, but you get the sense that you need very different architectures for different use cases.
diegom
So it seems to me that GPU development has mostly been driven by gaming (correct me if I'm wrong though). How is it that people have been able to so effectively use hardware that is specialized for rendering images for other things like deep learning?
Alex
@diegom I suppose both graphics tasks and ML tasks require the same computations: vector operations. Seems like a lot of the graphics pipeline comes down to multiplying matrices. Since our parametric models are of the form f(Wx+b) for some nonlinear f, both forward and backward passes are just matrix operations, so we should be able to send these operations off to the GPU.
emmaloool
@diegom Companies that produce GPUs no longer really focus only on graphics functionality. In fact, the industry is much more focused on making shader cores to follow the unified shader model, which enables both compute and graphics workloads to be run. These compute workloads are highly parallelizable, using basically the same threading models to partition computation like in graphics.
So really, it's thanks to graphics and GPU engineers and the trends of the industry that the fixed-functionality of GPUs has been overcome to enable more general-purpose computation.
Its funny how different in size GPUs can be even though they all accomplish the same task.
@mdesoto Same task, different speed! Very, very roughly speaking, A13 does about 1 trillion operations per second (for the whole chip, including CPU+GPU), whereas a GeForce 3090 RTX does about 36 trillion operations per second. The former has about 8.5 billion transistors; the latter about 28.3 billion. Also, totally different power budget. Not an apples-to-apples comparison at all, but you get the sense that you need very different architectures for different use cases.
So it seems to me that GPU development has mostly been driven by gaming (correct me if I'm wrong though). How is it that people have been able to so effectively use hardware that is specialized for rendering images for other things like deep learning?
@diegom I suppose both graphics tasks and ML tasks require the same computations: vector operations. Seems like a lot of the graphics pipeline comes down to multiplying matrices. Since our parametric models are of the form f(Wx+b) for some nonlinear f, both forward and backward passes are just matrix operations, so we should be able to send these operations off to the GPU.
@diegom Companies that produce GPUs no longer really focus only on graphics functionality. In fact, the industry is much more focused on making shader cores to follow the unified shader model, which enables both compute and graphics workloads to be run. These compute workloads are highly parallelizable, using basically the same threading models to partition computation like in graphics.
So really, it's thanks to graphics and GPU engineers and the trends of the industry that the fixed-functionality of GPUs has been overcome to enable more general-purpose computation.